[Pacemaker] Required resources for stateful clones (Serge Dubrouski)
Daniel Bozeman
daniel.bozeman at americanroamer.com
Fri May 20 14:14:45 UTC 2011
# postgresql.conf on both servers
hot_standby = true
wal_level = hot_standby
max_wal_senders = # not 0
# recovery.conf on both servers
standby_mode = 'on'
primary_conninfo = 'host=[hostname]'
trigger_file = '/tmp/trigger'
Then all you have to do is rename recovery.conf on one server to recovery.done to make it master. Start postgres and streaming replication will begin (once you rebase). Should the master go down, have Pacemaker touch the trigger file on the standby server and restart postgres. This will move recovery.conf to recovery.done and the standby will be master.
When you are ready to recover the old master, move recovery.done to recovery.conf, rebase, and start postgres. Pgpool has C and SQL functions you can install that will allow you to perform online recovery (rebasing the standby without shutting down the master).
Here is an example rebase script written in Ruby. It is easy to find them written in shell script as well, but this script should be very easy to translate as it is nothing but system calls.
#!/usr/bin/env ruby
# performs rebase from current machine
# to machine "to_host"
# remote host
host = 'ha2'
print 'recovering...'
system("ssh -T #{host} /etc/init.d/postgresql stop")
system("psql -U postgres -c \"SELECT pg_start_backup('backup', true)\" postgres")
system("rsync --exclude server.key --exclude server.crt --exclude pg_xlog --exclude recovery.conf --exclude recovery.done --exclude postmaster.pid --exclude postgresql.pid --exclude \"pg_log/*\" --exclude \"pg_hba.conf\" --exclude \"postgresql.conf\" --recursive --times --compress --perms --links --delete /var/lib/postgresql/9.0/main/ postgres@#{host}:/var/lib/postgresql/9.0/main/")
system("psql -U postgres -c \"SELECT pg_stop_backup()\" postgres")
system("rsync -av /var/lib/postgresql/9.0/main/pg_xlog/ postgres@#{host}:/var/lib/postgresql/9.0/main/pg_xlog/")
system("ssh -T #{host} mv /var/lib/postgresql/9.0/main/recovery.done /var/lib/postgresql/9.0/main/recovery.conf")
system("ssh -T #{host} /etc/init.d/postgresql start")
Is this a postgres mailing list or something? ;)
See one inline comment below.
Take care,
Daniel
On May 20, 2011, at 9:04 AM, Eamon Roque wrote:
>
> > On Fri, May 20, 2011 at 7:09 AM, Eamon Roque <Eamon.Roque at lex-com.net>wrote:
> >
> > >
> > > >
> > > > On Fri, May 20, 2011 at 3:42 AM, Eamon Roque <Eamon.Roque at lex-com.net
> > > >wrote:
> > > >
> > > > > Hi,
> > > > >
> > > > >
> > > > > >> On Thu, May 19, 2011 at 5:05 AM, Eamon Roque <
> > > Eamon.Roque at lex-com.net
> > > > > >wrote:
> > > > >
> > > > > >> Hi,
> > > > > >>
> > > > > >> I've put together a cluster of two nodes running a databank without
> > > > > shared
> > > > > >> storage. Both nodes replicate data between them, which is taken care
> > > of
> > > > > by
> > > > > >> the databank itself.
> > > > > >>
> > > > > >> I have a resource for the databank and ip. I then created a stateful
> > > > > clone
> > > > > >> from the databank resource. I created colocation rules joining the
> > > > > >> databank-ms-clone and ip:
> > > > > >>
> > > > > >> node pgsqltest1
> > > > > >> node pgsqltest2
> > > > > >> primitive Postgres-IP ocf:heartbeat:IPaddr2 \
> > > > > >> params ip="10.19.57.234" cidr_netmask="32" \
> > > > > >> op monitor interval="30s" \
> > > > > >> meta is-managed="false"
> > > > > >> primitive resPostgres ocf:heartbeat:pgsql \
> > > > > >> params pgctl="/opt/PostgreSQL/9.0/bin/pg_ctl"
> > > > > >>pgdata="/opt/PostgreSQL/9.0/data" psql="/opt/PostgreSQL/9.0/bin/psql"
> > > > > >> pgdba="postgres" \
> > > > > >> op monitor interval="1min" \
> > > > > >> meta is-managed="false"
> > > > > >> ms msPostgres resPostgres \
> > > > > >> meta master-max="1" master-node-max="1" clone-max="2"
> > > > > >> clone-node-max="1" notify="true" target-role="started"
> > > > > >> colocation colPostgres inf: Postgres-IP msPostgres:Master
> > > > > >> order ordPostgres inf: msPostgres:promote Postgres-IP:start
> > > > > >> property $id="cib-bootstrap-options" \
> > > > > >> dc-version="1.1.2-2e096a41a5f9e184a1c1537c82c6da1093698eb5"
> > > \
> > > > > >> cluster-infrastructure="openais" \
> > > > > >> expected-quorum-votes="2" \
> > > > > >> stonith-enabled="false" \
> > > > > >> no-quorum-policy="ignore" \
> > > > > >> last-lrm-refresh="1302707146"
> > > > > >> rsc_defaults $id="rsc-options" \
> > > > > >> resource-stickiness="200"
> > > > > >> op_defaults $id="op_defaults-options" \
> > > > > >> record-pending="false"
> > > > > >>
> > > > > >> The normal postgres agent doesn't support this functionality, but
> > > I've
> > > > > put
> > > > > >> together my own using the mysql agent as a model. Before running the
> > > > > script
> > > > > >> through ocf-tester, I unmanage the postgres resource.
> > > > > >>
> > > > >
> > > > > > Could you show how you implemented promote/demote for pgsql?
> > > > >
> > > > > Sure, let's start with the ultra-simple "promote" function:
> > > > >
> > > > > #
> > > > > # These variables are higher up in the file, but they will probably
> > > help
> > > > > with understanding the error of
> > > > > # my ways.
> > > > >
> > > > > CRM_MASTER="${HA_SBIN_DIR}/crm_master"
> > > > > ATTRD_UPDATER="${HA_SBIN_DIR}/attrd_updater"
> > > > >
> > > > > pgsql_promote() {
> > > > > local output
> > > > > local rc
> > > > > local CHECK_PG_SQL
> > > > > local COMPLETE_STANDBY_QUERY
> > > > > local PROMOTE_SCORE_HIGH
> > > > > local MOD_PSQL_M_FORMAT
> > > > >
> > > > >
> > > > > PROMOTE_SCORE_HIGH=1000
> > > > > CHECK_PG_SQL="SELECT pg_is_in_recovery()"
> > > > > MOD_PSQL_M_FORMAT="$OCF_RESKEY_psql -Atc"
> > > > > COMPLETE_STANDBY_QUERY="$MOD_PSQL_M_FORMAT \"$CHECK_PG_SQL\""
> > > > >
> > > > > output=$(su - $OCF_RESKEY_pgdba -c "$COMPLETE_STANDBY_QUERY"
> > > 2>&1)
> > > > > echo $output
> > > > >
> > > > > rc=$?
> > > > >
> > > > > case $output in
> > > > > f)
> > > > > ocf_log debug "PostgreSQL Node is running in
> > > Master
> > > > > mode..."
> > > > > return $OCF_RUNNING_MASTER
> > > > > ;;
> > > > >
> > > > > t)
> > > > > ocf_log debug "PostgreSQL Node is in
> > > Hot_Standby
> > > > > mode..."
> > > > > return $OCF_SUCCESS
> > > > > ;;
> > > > >
> > > > > *)
> > > > > ocf_log err "Critical error in $CHECK_PG_SQL:
> > > > > $output"
> > > > > return $OCF_ERR_GENERIC
> > > > > ;;
> > > > > esac
> > > > >
> > > > > #
> > > > > # "Real" promotion is handled here.
> > > > > # The trigger file is created and we check for "recovery.conf" on the
> > > host.
> > > > > # If we can't find it, then the file will be copied from the HA-Config
> > > into
> > > > > postgres' data folder.
> > > > > #
> > > > >
> > > > > if ! touch $OCF_RESKEY_trigger_file; then
> > > > > ocf_log err "$OCF_RESKEY_trigger_file could not be created!"
> > > > > return $OCF_ERR_GENERIC
> > > > > fi
> > > > >
> > > > > if [ ! -f $OCF_RESKEY_recovery_conf ]; then
> > > > > ocf_log err "$OCF_RESKEY_recovery_conf doesn't exist!"
> > > > > cp $OCF_RESKEY_recovery_conf_ersatz $OCF_RESKEY_pgdata
> > > > > return $OCF_SUCCESS
> > > > > fi
> > > >
> > > >
> > > > Why do you need this? As far as I know when you switch standby database
> > > to
> > > > primary using trigger file recovery.conf gets renamed to recovery.done.
> > > If
> > > > you rename it back DB will be put into standby mode after restart.We are
> > > > talking about streaming replication, right?
> > > >
> > > >
> > > Right. The order is wrong. According to the Binary Replication tutorial on
> > > the postgres wiki, when I perform a failover with a trigger file, it wants
> > > to find a "recovery.conf", which it then processes (checking the archive for
> > > missing updates etc.) and renames (after noticing the trigger file).
> > >
> > > I assumed that this would work in exactly the same way with Streaming
> > > Replication.
> > >
> > > Am I wrong?
> >
> >
> > I think so. You have to have recovery.conf when you start your standby, not
> > master. Actually instance that has recovery.conf always tries to start as
> > standby. You have to have master's IP address there and path to archived log
> > files.
> >
> >
>
> So the failover behavior in binary replication and streaming replication is different? Or is the wiki entry just antiquated?
"Binary replication" simply refers to streaming replication between two (binary) servers. This is a bit misleading, as it may lead you to believe it is a WAL-shipping or shared storage replication method, but it is not. The information I have provided above should be all you need to set up streaming/binary replication (other than properly configuring your pg_hba.conf
>
> > >
> > >
> > > > >
> > > > >
> > > > > # If both file exist or can be created, then the failover fun can
> > > start.
> > > > >
> > > > > ocf_log info "$OCF_RESKEY_trigger_file was created."
> > > > > ocf_log info "$OCF_RESKEY_recovery_conf exists and can be copied to the
> > > > > correct location."
> > > > >
> > > > > # Sometimes, the master needs a bit of time to take the reins. So...
> > > > >
> > > > > while :
> > > > > do
> > > > > pgsql_monitor warn
> > > > > rc=$?
> > > > >
> > > > > if [ $rc -eq $OCF_RUNNING_MASTER ]; then
> > > > > break;
> > > > > fi
> > > > >
> > > > > ocf_log debug "Postgres Server could not be promoted. Please
> > > > > wait..."
> > > > >
> > > > > sleep 1
> > > > >
> > > > > done
> > > > >
> > > > > ocf_log info "Postgres Server has been promoted. Please check on the
> > > > > previous master."
> > > > >
> > > > > #################################
> > > > > #Attributes Update: #
> > > > > #################################
> > > > >
> > > > > $ATTRD_UPDATER -n $PGSQL_STATUS_NAME -v \"PRI\" || exit $(echo "Eh!
> > > > > Attrd_updater is not working!")
> > > > >
> > > > > #############################################
> > > > > # Resource stickiness pumped up to 1000 : #
> > > > > #############################################
> > > > >
> > > > > $CRM_MASTER -v $PROMOTE_WERT_HOCH || exit $(echo "crm_master could not
> > > > > change the Master's status!")
> > > > >
> > > > > ############
> > > > > # Success! #
> > > > > ############
> > > > >
> > > > > return $OCF_SUCCESS
> > > > >
> > > > > }
> > > > >
> > > > >
> > > > >
> > > >
> > >
> > ######################################################################################################
> > > > >
> > > > > Thanks!
> > > > >
> > > > >
> > > > And what about demote? Switching standby into primary using trigger files
> > > > changes TIMELINE in the DB and that invalidates all other standby
> > > databases
> > > > as well as previous master database. After that you have to restore them
> > > > from a fresh backup made on new master. This particular behavior stopped
> > > me
> > > > from implementing Master/Slave functionality in pgsql RA so far.
> > > >
> > > > BTW, why pgsql is set to is-managed="false" in your configuration.With
> > > this
> > > > setting cluster will keep monitoring it but won't take any other actions
> > > > AFAIK.
> > >
> > > Demote? Well, seeing as neither promote nor demote actually worked for me,
> > > I thought I would start small.
> > >
> >
> > It doesn't work because you have it in unmanaged state I think.
> >
>
> I'm using the ocf-tester utility to test the agent. Won't there be a conflict if I try and have the cluster manage the resources and then try and wrest it's control away with my own testing agent?
>
>
> >
> > >
> > > As far as the trigger file switching goes, you're of course completely
> > > right. This behavior isn't really a big deal in my environment, as it's
> > > meant as more of test and we want to bring back the demoted servers up
> > > manually, but I can see that it would cause a lot of problems in a more
> >
> >
> > That means that demote operation should stop master server which isn't the
> > best behavior IMHO.
> >
>
> I don't disagree. This was the policy that was "agreed" upon, so it's more of a political issue, really.
>
> Would you prefer putting it into RO mode?
>
> >
> >
> > > complex environment. When I tested the failover functionality without
> > > pacemaker, I have to perform a fresh backup even if I waited less than 30s
> > > to bring the old master back up as a standby.
> > >
> > > I guess that with 9.1 this will be easier...
> > >
> > > I unmanaged the resources so that my test agent would handle them. Is this
> > > incorrect?
> > >
> >
> > Again I think you are wrong. In this mode pacemaker won't call your RA to
> > promote/demote or failover your resource.
> >
> >
> > >
> > >
> > > >
> > > >
> > > > ?amon
> > > > >
> > > > >
> > > > >
> > > > > >> Unfortunately, promote/demote doesn't work. ocf-tester tries to use
> > > the
> > > > > >> "crm_attribute -N pgsql1 -n master-pgrql-replication-agent -l reboot
> > > -v
> > > > > >> 100", but the (unmanaged) resources don't accept the score change.
> > > > > >>
> > > > > >> I'm pretty sure that I just need to be hit with a clue stick and
> > > would
> > > > > be
> > > > > >> grateful for any help.
> > > > > >>
> > > > > >> Thanks,
> > > > > >>
> > > > > >> ?amon
> > > > > >>
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Serge Dubrouski.
> > > > > _______________________________________________
> > > > > Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> > > > > http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> > > > >
> > > > > Project Home: http://www.clusterlabs.org
> > > > > Getting started:
> > > http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> > > > > Bugs:
> > > > >
> > > http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
> > >
> > > > >
> > > > >
> > > >
> > > >
> > > > --
> > > > Serge Dubrouski.
> > > > -------------- next part --------------
> > > > An HTML attachment was scrubbed...
> > > > URL: <http://oss.clusterlabs.org/pipermail/pacemaker/attachments/
> > > > 20110520/e1f26230/attachment.html>
> > > >
> > > > ------------------------------
> > >
> > > >
> > > > _______________________________________________
> > > > Pacemaker mailing list
> > > > Pacemaker at oss.clusterlabs.org
> > > >
> > > http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> > > >
> > > >
> > > > End of Pacemaker Digest, Vol 42, Issue 53
> > > > *****************************************
> > >
> > > _______________________________________________
> > > Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> > > http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> > >
> > > Project Home: http://www.clusterlabs.org
> > > Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> > > Bugs:
> > > http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
> > >
> > >
> >
> >
> > --
> > Serge Dubrouski.
> > -------------- next part --------------
> > An HTML attachment was scrubbed...
> > URL: <http://oss.clusterlabs.org/pipermail/pacemaker/attachments/
> > 20110520/19777245/attachment.html>
> >
> > ------------------------------
> >
> > _______________________________________________
> > Pacemaker mailing list
> > Pacemaker at oss.clusterlabs.org
> > http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> >
> >
> > End of Pacemaker Digest, Vol 42, Issue 55
> > *****************************************
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
>
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
Daniel Bozeman
American Roamer
Systems Administrator
daniel.bozeman at americanroamer.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.clusterlabs.org/pipermail/pacemaker/attachments/20110520/01a5fcb6/attachment.htm>
More information about the Pacemaker
mailing list